Re: [RFC] More jump threading restrictions in the presence of loops.

[Aldy Hernandez <aldyh@redhat.com>]

PING.

Note, that there are no PRs and nothing really dependent on this
patch.  This has just been suggested as the right thing to do wrt
loops.

This patch fixes 6 XFAILs in our testsuite and has the added
side-effect of fixing the aarch64 bootstrap problem (though the
problem in the uninit code is still there).

This is a fundamental change in what we've traditionally allowed for
jump threading, but I think it's a good thing.  It also paves the way
for combining the validity models for both the forward and the
backward threaders.

I am happy to field the PRs this may bring about, since every change
in the cost model has us questioning whether we should or shouldn't
thread a path.  But I may need some help from y'all if there's a
missing thread that causes a regression in some other pass.  That
being said, most of the issues that have come with the threader
changes haven't been because we thread less, but because we thread
more-- so perhaps restricting things is a good thing ;-).

Aldy

On Wed, Oct 6, 2021 at 12:22 PM Aldy Hernandez <aldyh@redhat.com> wrote:
>
> [Here go the bits by Richi, tested on x86-64 Linux, and ongoing tests
> on aarch64 and ppc64le.]
>
> There is a lot of fall-out from this patch, as there were many threading
> tests that assumed the restrictions introduced by this patch were valid.
> Some tests have merely shifted the threading to after loop
> optimizations, but others ended up with no threading opportunities at
> all.  Surprisingly some tests ended up with more total threads.  It was
> a crapshoot all around.
>
> On a postive note, there are 6 tests that no longer XFAIL, and one
> guality test which now passes.
>
> I would appreciate someone reviewing the test changes.  I am unsure
> whether some of the tests should be removed, XFAILed, or some other
> thing.
>
> I felt a bit queasy about such a fundamental change wrt threading, so I
> ran it through my callgrind test harness (.ii files from a bootstrap).
> There was no change in overall compilation, DOM, or the VRP threaders.
>
> However, there was a slight increase of 1.63% in the backward threader.
> I'm pretty sure we could reduce this if we incorporated the restrictions
> into their profitability code.  This way we could stop the search when
> we ran into one of these restrictions.  Not sure it's worth it at this
> point.
>
> Note, that this ad-hoc testing is not meant to replace a more thorough
> SPEC, etc test.
>
> Tested on x86-64 Linux.
>
> OK pending tests on aarch64 and ppc64le?
>
> Co-authored-by: Richard Biener <rguenther@suse.de>